Google rolls out Gemini 4 Carbon, a cybersecurity‑focused frontier model

Can an AI model reliably help security teams hunt for bugs, verify exploits and generate patches without drowning developers in false leads? Google says its newest Gemini 4 Carbon model is built for exactly that, targeting multi‑step engineering problems that can take days for humans to solve. Carbon is positioned as a long‑horizon, agentic system that persists through dead ends, aiming to turn weeks of manual triage into minutes of automated insight.

The announcement follows Gemini 4 Argon, released only weeks earlier, and highlights a rapid escalation in capability. By coupling stronger coding performance with a dedicated cybersecurity focus, Carbon promises to shift how organizations approach vulnerability management, moving from reactive patching toward proactive, AI‑assisted defense.

Source: Google Gemini Blog: Gemini 4 Carbon: our next frontier model ↗

Google reports that Carbon tops the v5.0 Artificial Analysis Intelligence Index and leads external benchmarks: 88.2% on SEC‑bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE‑bench for automated patching. Internally, the model delivers more than a two‑fold uplift in secure‑code generation compared with Argon, and live teams saw an 80% drop in high‑severity misalignment alerts.

Safety claims are backed by a 0.0% success rate on the Gray Swan Indirect Prompt Injection benchmark, improving on Argon’s 0.7% rate. Pricing for the introductory phase matches Argon at $2 per million input tokens and $10 per million output tokens, with cached inputs heavily discounted.

Source: Google Gemini Blog: Gemini 4 Carbon: our next frontier model ↗

Carbon is not yet open to the public. Google is rolling it out first to a limited cohort of cyber defenders through its Fairwind Program and is participating in the U.S. government’s voluntary pre‑release testing process. The model’s safety guardrails are still being refined, and Google notes that pricing will rise to $4 per million input tokens and $20 per million output tokens after the introductory period.

Because the model is evaluated without guardrails on external benchmarks, real‑world performance may differ when safety layers are active. Organizations must also consider integration effort, as Carbon’s agentic workflow expects multi‑step planning and persistent context that may not fit all existing toolchains.

If you’re a security professional or developer interested in early access, apply to Google’s Fairwind Program or watch for announcements about expanded partner invitations. Meanwhile, experiment with Gemini 4 Argon, which is already available to developers, to gauge how Google’s multimodal agents fit your workflow before Carbon’s broader release.

Stay tuned to Google’s AI blog for benchmark updates and safety reports, and begin mapping your vulnerability‑to‑patch pipeline to identify where a long‑horizon agent could add value. Early pilots will give you a head‑start on integrating AI‑driven security once Carbon becomes generally available.